berthPorts your agents and your terminal agree on

Every project gets a permanent, decodable block of ports. A ledger records who holds what, a reconciler checks it against what is really listening, and every Claude Code session is told its ports before it binds anything. Advisory by design; it never kills.

Getting started

Claude Code and agents

GitHub

The number is the rule
port = 10000 + 1000·P + 100·W + R. Reading a port is decoding it: 13204 is project 3, worktree 2, smtp. No allocation table to remember.
Truth, not guesses
lsof, netstat, Docker compose labels and the Claude session marker in a process's environment attribute every listener, including containers behind Colima or Docker Desktop.
Built for agent sessions
A Claude Code plugin injects each session's ports at start, puts berth on its PATH, exposes MCP tools and ships skills. Commands with teeth are fenced off from agents.
A map you can read
One row per project, live legacy ports as numbered cells, eight states with an advisory each, and a drawer that shows the evidence for any port.

Port 13204 decoded: base 10000, P=3 breach-resolve, W=2 second worktree, R=04 smtp

Sixty seconds

bash

npm install -g @amiable-dev/berth
berth init                                   # a starting policy in ~/.config/berth/policy.toml
cd ~/projects/my-app && berth project add .  # permanent project number, ports found in your configs
eval "$(berth env --shell)"                  # PORT, API_PORT, DB_PORT … for this checkout
berth check                                  # what is listening, who holds it, what needs attention

With Claude Code, install the plugin instead and let the agent register its own repository:

bash

claude plugin marketplace add amiable-dev/berth
claude plugin install berth@berth

berthPorts your agents and your terminal agree on

The number is the rule

Truth, not guesses

Built for agent sessions

A map you can read

Sixty seconds ​

Sixty seconds